Numerous applications in clinical medicine and forensic sciences depend on reliable data concerning the frequencies of human leukocyte antigen (HLA) genes and haplotypes. Assuming a Hardy-Weinberg equilibrium of the underlying population, these frequencies can be estimated from phenotype data using an expectation-maximization-algorithm also known under the name "gene counting." We have refined this algorithm in order to cope with the heterogeneous resolution of HLA phenotypes frequently occurring in large datasets due to the structure of the HLA nomenclature. This was a prerequisite to analyze a set of 13,386 blood donors contributed by over 40 blood banks who were tested for HLA-DR when they volunteered to become marrow donors. This data set is still unique in the German national donor registry because their HLA-DR-typing was not biased by patient oriented searches or other strategies for selective typing. As a consequence of the size of the sample, the frequency estimates for the genes and the two- and three-locus haplotypes of HLA-A, HLA-B, and HLA-DR are of unprecedented precision and allow interesting projections concerning the efficiency and economic aspects of the development of a large donor registry in Germany.